Boosting autonomous process design and intensification with formalized domain knowledge. (January 2023)

• Record Type: Journal Article
• Authors: Seidenberg, J. Raphael; Khan, Ahmad A.; Lapkin, Alexei A.
• Is Part Of: Computers & chemical engineering. Volume 169(2023)
• Publication Date: 2023-01
• Subjects: RL reinforcement learning -- HRL hierarchical reinforcement learning -- HL higher-level -- LL lower-level -- s state -- a action -- π action policy -- r reward.; Process synthesis -- Process intensification -- Ontology -- Knowledge graph -- Reinforcement learning -- Machine learning; Chemical engineering -- Data processing -- Periodicals; 660.0285
• DOI: 10.1016/j.compchemeng.2022.108097
• Languages: English
• ISSNs: 0098-1354

Abstract

Abstract: Conceptual process design deals with searching for optimal process flowsheets in a large design space. Effective approaches benefit from sensible search space restrictions, commonly carried out by a knowledgeable expert in the form of a superstructure optimization or heuristic rules. To achieve the goal of autonomous process design, knowledge has to be formalized in a machine-readable format. This contribution aims to incorporate an ontological representation of fundamental process knowledge to empower general-purpose design procedures while respecting problem-specific variability. Specifically, the presented framework leverages an ontology to express declarative knowledge (what-is) of processes, phenomena and design tasks to set up the search space and boost a hierarchical reinforcement learning agent which learns the required procedural knowledge (how-to) in order to find an optimal solution. The work is applied in a case study of an intensified steam methane reforming process. Results show that the automated treatment of domain knowledge allows for dynamic search space reduction and achieves better computational efficiency and solution quality, highlighting its potential in autonomous process design approaches. Highlights: A framework is developed for autonomous process design and intensification Process design ontology and knowledge graph to specify the space of designs Hierarchical reinforcement learning to discover optimal designs within feasible space A case study shows value of combined knowledge-based, data-based and numerical tools General-purpose procedure is proposed for end-to-end autonomous design